Which means

if someone comes along and is an attacker and they know the hash of that resource, which is not a secret, so hashes are just something that everyone can calculate, they could just probe the cost cache and see like, hey, do you have the resource ABC123? And if it happens to be that particular resource, logged in user logo, they could determine that you are likely a customer of the bank because you have that resource in the cache.

You must have logged in at one point in time.

So this is why we said, like, wait, this needs to not happen.

We need to make this particular attack impossible.

Which is why we came up with what we call a proposal that we call the public hash list.

And the public hash list has the idea of collecting resources that are known to be popular.

So have it just go booga dooga dooga download 3.9. gigabytes of crap or whatever but then I'm using it everywhere I'm using it on any implementation I have a model choice on any website and I only had to download it once didn't you straight up say like that's the main motivation for this time like essentially like making web LLMs faster, but not only faster, but more shareable?

And those

AI applications try to load a model like, let's say, the Whisper audio transcription model.

And you use two different apps that would depend on the same model.

But of course, each of them has to download the model again.

And it starts to become prohibitive relatively quickly.

which is why we have seen a lot of web AI demos, but we've not seen that many actual production roll-ons with AI.

People might remember the last time I was on the show, I forgot when, a year ago maybe, when we talked about web AI, but in the context of built-in AI.

What would make you happy if people played with it? Where can they get feedback? What kind of feedback are you looking for?